A learning algorithm for improving generalization ability of multi layered neural network for pattern recognition problem

This paper proposes a new learning algorithm for improving generalization ability of multilayered neural network for pattern recognition problem. We discuss relationships between the internal representation and the generalization ability for pattern recognition problem, and show two important characteristics of the internal representation for the generalization ability. Based on the above discussion, we propose a new learning algorithm for improving generalization ability which makes changes of output units for input units small as possible. The proposed algorithm is applied to printed numeral fonts recognition problem and we show the effectiveness of this algorithm compared with other algorithms